Step two. When it became clear the problem wasn't the VPN

Once the VPN was more or less settled — several servers, masking, chains — a new question came up: what's next?

And here's an important aside.

Since 2020, when I was in the jungle, I'd had a channel and a chat. The audience kept growing — and along with it, more and more time went into communication.

Messages. Questions. Personal conversations.

At some point, time simply ran out. I started cutting things back: leaving chats, closing discussions, limiting the format. But the people kept coming.

I saw the same thing happening to others — anyone running social accounts or working with an audience. Communication is necessary. But it eats up all your time.

And then it became obvious: if it's a recurring problem, it needs to be automated. First — for myself. Then maybe as a separate product.

The idea was simple: an auto-responder in Telegram that talks on my behalf. Something that looks like a normal conversation. With its own style. Tailored to the specific person on the other end. Without obvious signs that the replies aren't human.

Not for the trick. For the time it saves.

The approach was the same as with the VPN. I didn't sit down to learn development. I just typed in the terminal what I wanted to get.

The first prototype appeared. Raw. I started testing it. Sometimes I forgot to turn it off — and people would stumble onto it and start talking.

It was interesting to watch how many messages they'd send and whether they'd ever realize it wasn't me replying.

Bit by bit the auto-responder turned into a normal tool: you could choose where it was active, set the style, control the logic.

In parallel I was experimenting with models. Tried local ones. It quickly became clear: with an 8 GB GPU, no miracles were happening.

But I did get one key thing very clearly: the model, for all its "smarts," is extremely forgetful. It constantly needs reminders of who it is, where it is, and why.

That applies to everything: working in the terminal, auto-responders, any kind of automation.

At some point an unexpected stress test showed up. People started writing me harsh messages demanding money. I turned the auto-responder on and just watched it carry the conversation.

That turned out to be more useful than any test. It immediately showed where it was too soft, where it started making excuses, and where it actually held the line better than I would have.

After that I rebuilt the response logic for those situations. And finally stopped treating the whole thing as a toy.

It was already a working tool that took load off me in places where a human only gets in the way.

I never took the project to a commercial version. The version for myself is sitting there — as a working sandbox.

And right there it hit me: this is how projects actually get built. First — a simple version for yourself. Then — expansion. Then — limits. Then — understanding what comes next.

But then I got pulled away sharply. A real project showed up — with orders, clients, delivery, and dispatchers.

And that turned out to be even more interesting.